def TpuTrain():
loop_result = tpu_training_loop.repeat(
self._train_steps_per_loop,
TpuTrainStep,
inputs=[],
name='train_loop')
return loop_result
def TpuEval():
loop_result = tpu_training_loop.repeat(
self._steps_per_loop,
TpuEvalStep,
inputs=self._eval_metrics.initial_values,
name='eval_loop')
# Final metrics are the avg across self._steps_per_loop steps.
return self._eval_metrics.FinalizeMetrics(loop_result)
def eval_loop(self):
per_replica_eval_batch_size = self.eval_batch_size // self.num_replicas
tf.get_variable_scope().reuse_variables()
predictions = tf.zeros([self.eval_steps, per_replica_eval_batch_size, 2])
_, predictions = training_loop.repeat(
int(self.eval_steps), self.eval_step, [tf.constant(0), predictions])
with tf.control_dependencies([tpu_ops.outfeed_enqueue_tuple([predictions])
]):
return tf.no_op()
def train_eval_loop():
return training_loop.repeat(self.hparams.max_train_epochs,
train_eval_step, [])
def eval_loop():
if self.eval_steps > 0:
return training_loop.repeat(self.eval_steps, tpu_eval_step, [])
else:
return tf.no_op()
def train_loop():
return training_loop.repeat(self.iterations, tpu_train_step,
[_INITIAL_LOSS])
def TrainAndDecodeLoop():
tpu_training_loop.repeat(
self.num_epochs_per_session_run, TrainAndDecode, inputs=[])
def DecodeLoopFn():
return tpu_training_loop.repeat(
self._decode_steps_per_loop, _DecodeStep, inputs=[])
def train_eval_loop():
return training_loop.repeat(self.max_train_iterations,
train_eval_step)
def train_loop():
return training_loop.repeat(self.iterations_per_loop, train_step,
tf.constant(0))
def eval_loop(self):
tf.get_variable_scope().reuse_variables()
return training_loop.repeat(int(self.eval_steps), self.eval_step)
